PaperlayerView exact revision on arXiv

Source-linked AI summary

Scale- and Context-Aware Convolutional Non-intrusive Load Monitoring

Kunjin Chen, Yu Zhang, Qin Wang, Jun Hu, Hang Fan, Jinliang He

arXiv:1911.07183v1eess.SPcs.LGstat.ML

TL;DR

NILM estimates appliance-level consumption from a household’s aggregate electricity signal, while CNN-based models do not fully exploit scale variation or context dependencies. The paper proposes SCANet, combining scale- and context-aware features with adversarial loss and on-state augmentation. It reports lower MAEs and SAEs than SGN across appliances, with especially large average improvements on REDD.

  • Problem

    CNN-based NILM models fail to exploit varied appliance time scales and contextual dependencies in consumption patterns.

    Source: §I · p. 1

  • Method

    SCANet combines scale- and context-aware CNN components with self-attention, adversarial loss, and on-state augmentation.

    Source: §I · p. 1, 2) Context-aware Feature Integration · p. 3

  • Results

    22.39% and 28.60% improvements in averaged MAEs and SAEs, respectively, are reported for REDD, with lower MAEs and SAEs than SGN for all appliances.

    Source: B. Experiment Results · p. 7

  • Takeaways & Limitations

    The authors conclude that the scale- and context-aware modules function as expected and contribute to improved disaggregation accuracy.

    Source: §V · p. 10

Abstract

from arXiv · show

Non-intrusive load monitoring addresses the challenging task of decomposing the aggregate signal of a household's electricity consumption into appliance-level data without installing dedicated meters. By detecting load malfunction and recommending energy reduction programs, cost-effective non-intrusive load monitoring provides intelligent demand-side management for utilities and end users. In this paper, we boost the accuracy of energy disaggregation with a novel neural network structure named scale- and context-aware network, which exploits multi-scale features and contextual information. Specifically, we develop a multi-branch architecture with multiple receptive field sizes and branch-wise gates that connect the branches in the sub-networks. We build a self-attention module to facilitate the integration of global context, and we incorporate an adversarial loss and on-state augmentation to further improve the model's performance. Extensive simulation results tested on open datasets corroborate the merits of the proposed approach, which significantly outperforms state-of-the-art methods.

I. INTRODUCTION

NILM estimates appliance-level power from a household’s aggregate consumption, and appliance-level feedback can help end users optimize energy behavior. The paper addresses scale variation and missing contextual information in CNN-based NILM with SCANet.

I. INTRODUCTION

  • NILM estimates a specific appliance’s power demand from a household’s aggregate consumption measured by a single meter.The paper also calls this task energy disaggregation; appliance-level feedback may help end users optimize energy behavior, with up to 12% residential energy savings estimated from such feedback.Source: §I · p. 1

1) Neural Non-intrusive Load Monitoring:

  • Supervised NILM uses individual-appliance consumption data for training, whereas unsupervised methods use aggregate power data.Unsupervised approaches discussed include HMMs, FHMMs, and event-detection and clustering methods.Source: §I · p. 1

I. INTRODUCTION

  • CNN-based NILM models do not explicitly capture varied time scales or context dependencies, which the authors say yields false-positive and false-negative disaggregation errors.The paper cites appliance patterns at different scales and behavioral sequences such as a dryer operating after a washer as examples.Source: §I · p. 1
  • SCANet is proposed as a scale- and context-aware CNN structure that improves disaggregation results for multiple appliances.The paper also reports that adversarial loss or on-state augmentation can improve accuracy and generalizability.Source: §I · p. 1

2) Multi-scale Features in CNNs:

  • SCANet uses parallel dilated-convolution branches and gating signals from the on/off classification network to selectively retain regression features.This differs from TridentNet’s approach of separately training branches for specified scale ranges.Source: 2) Multi-scale Features in CNNs · p. 2

3) Self-attention Mechanism:

The paper introduces self-attention as a way to integrate information from across a sequence and describes an adversarial approach for NILM. It formulates the task as recovering appliance power sequences from an aggregate household signal.

3) Self-attention Mechanism:

  • Self-attention calculates a sequence position’s output by attending to all positions, allowing global context to inform the representation.The paper adopts the self-attention module proposed by Zhang et al. [27].Source: 3) Self-attention Mechanism · p. 2

4) Generative Adversarial Networks:

  • Unlike prior NILM GAN work, the paper maps aggregate consumption directly to appliance-level consumption without producing latent representations.The preceding approach uses a disaggregator to produce latent representations followed by a generator that produces the appliance load sequence.Source: 4) Generative Adversarial Networks · p. 2

III. PROPOSED MODEL AND TRAINING TECHNIQUES

  • The proposed-model section introduces the NILM task, CNN-based models, SCANet building blocks, and training techniques..Source: §III · p. 2

A. Problem Formulation

  • The task is to recover appliance power-consumption sequences from the aggregate signal, which includes the target appliances, other appliances, and additive noise.The task illustration describes recovering a dishwasher’s signal from aggregate consumption.Source: A. Problem Formulation · p. 2, Figure 1 · p. 2

B. Model Design

The model uses an input window extended beyond the target output interval and two subnetworks to estimate appliance power and on/off state. Its design builds on an existing two-subnetwork CNN structure while adding scale and context awareness.

B. Model Design

  • The input aggregate sequence includes windows of length w on both sides of the target output interval to provide contextual information beyond that interval.The target output sequence estimates appliance consumption over length s.Source: B. Model Design · p. 2
  • The on/off subnetwork predicts an on-state probability sequence, which is used with the power subnetwork’s output to form the final model output.The output combines the two predictions by element-wise multiplication.Source: B. Model Design · p. 2
  • The model builds scale- and context-awareness modules on the structure of two sub-networks proposed in, addressing convolutional layers’ inability to explicitly extract features with different time scales and their lack of context.Source: B. Model Design · p. 3

1) Scale-aware Feature Extraction:

SCANet captures features at different time scales with parallel branches using different dilation rates and branch-wise gates. The gates connect the sub-networks and select regression feature maps at each scale.

1) Scale-aware Feature Extraction:

  • With the same number of layers and parameters, larger dilation rates respond to wider input time ranges, producing features at different time scales.The paper describes this as scale awareness added through parallel branches to the original network.Source: 1) Scale-aware Feature Extraction · p. 3
  • Branch-wise gates connect the two sub-networks and let the regression network retain important feature maps at different scales.Gating is performed separately for each dilation rate, enabling a combination of features across time scales.Source: 1) Scale-aware Feature Extraction · p. 3

2) Context-aware Feature Integration:

SCANet integrates global context through self-attention and adds an adversarial-loss training option using a critic. The self-attention output is combined with the local representation, while adversarial loss is intended to encourage more realistic output sequences, especially with relatively small training datasets.

2) Context-aware Feature Integration:

  • Self-attention lets each time step attend across the input sequence, then adds the resulting feature map to the local representation with a coefficient initialized at zero.The coefficient is updated during training, allowing the model to gradually learn dependencies in global context.Source: 2) Context-aware Feature Integration · p. 3
  • The two-subnetwork model combines mean squared error for overall disaggregation with binary cross-entropy for on/off-state classification.Source: 2) Context-aware Feature Integration · p. 3

1) Training with adversarial loss:

  • The adversarial-loss variant adds a critic and trains the model partially as a WGAN-GP, with the network output serving as the generator.The loss is intended to encourage more realistic output sequences, especially when training data is relatively limited.Source: 1) Training with adversarial loss · pp. 3–4

2) On-state augmentation:

On-state augmentation addresses variation in appliance power while an appliance is operating by offsetting both the input and output sequences during training. For fridge, the paper reports more accurate on-state power estimates after applying the augmentation.

2) On-state augmentation:

  • The augmentation targets variation in an appliance’s on-state power, replacing each output sequence with an offset version and applying the same offset to its corresponding input.The offset is sampled uniformly between appliance-specific minimum and maximum values, and augmentation is used during training.Source: 2) On-state augmentation · p. 4
  • After on-state augmentation, SCANet estimates fridge on-state power more accurately; biased on-state power estimates are a major source of its disaggregation error.Source: 2) On-state augmentation · p. 4

IV. RESULTS AND ANALYSIS

  • The results section presents datasets and implementation details alongside empirical analyses of SCANet’s advantages.Source: §IV · p. 4

A. Experiment Settings

Experiments use REDD and UK-DALE, with household-based training and test splits and appliance-specific disaggregation tasks. The implementation extends the SGN backbone and evaluates appliance-level error metrics across both datasets.

1) Datasets:

  • REDD covers three appliances and UK-DALE five, with dataset-specific results reported for both.REDD contains six US households; UK-DALE contains five UK households, with different sampling intervals and recording spans.Source: 1) Datasets · p. 4, Table I · p. 5, Table II · p. 5
  • REDD houses 2–6 train and house 1 tests; UK-DALE houses 1 and 5 train and house 2 tests.The reported appliances are fridge, dishwasher, and microwave for REDD, and those plus washing machine and kettle for UK-DALE; both datasets are normalized by 612.Source: 1) Datasets · p. 4

2) Implementation Details:

  • The SGN backbone has six convolutional layers and two fully connected layers per sub-network, with dataset-specific sequence and kernel sizes.SCANet adds branches with dilation rates 2 and 3 from the fourth convolutional layer and uses 64-channel self-attention modules.Source: 2) Implementation Details · pp. 4–5
  • Training uses Adam at learning rate 0.0001 for five epochs, while overlapping test-window outputs are averaged.Training samples use sliding windows; step sizes vary by dataset and appliance, and testing uses a step size of 2.Source: 2) Implementation Details · pp. 5–6
  • Off-state samples are randomly removed for selected appliances, and fridge models use on-state augmentation.The off-state retention probabilities vary by dataset and appliance; augmentation ranges also differ between REDD and UK-DALE.Source: 2) Implementation Details · p. 6
  • MAE and SAE evaluate each appliance; SAE compares predicted and ground-truth total power over time periods.With 1,200 periods, each period is approximately one hour for REDD and two hours for UK-DALE.Source: 2) Implementation Details · p. 6

B. Experiment Results

SCANet improves appliance disaggregation over SGN on both datasets, with especially large average gains on REDD. Visual analyses relate its behavior to edge-sensitive branch gating and self-attention, while regression features help track power-consumption trends.

B. Experiment Results

  • On REDD, SCANet achieves 22.39% and 28.60% improvements over SGN in averaged MAEs and SAEs, respectively; UK-DALE average improvements are also greater than 10%.Across all appliances, SCANet has lower MAEs and SAEs than SGN.Source: B. Experiment Results · p. 7
  • SCANet produces more accurate disaggregation than SGN, with fewer false positives and better power-level estimates in the illustrated cases.The examples include fridge on-state augmentation and microwave periods with successive short activations that SGN misses.Source: B. Experiment Results · p. 7
  • Across the microwave branches, features respond to rising and falling edges, while classification-branch gates suppress high regression activations.The resulting regression feature map is less activated overall.Source: B. Experiment Results · p. 7
  • Self-attention assigns high attention among rising edges, and its added feature map suppresses microwave on-state probability when the appliance is off.In the illustrated case, bypassing self-attention removes that suppression even though the appliance turns on shortly before and after.Source: B. Experiment Results · p. 8
  • For a microwave case, regression features respond to both edge directions and predict the power-consumption trend.This illustrates a role for the regression sub-network alongside classification.Source: B. Experiment Results · p. 8

C. Ablation Study and Discussion

The REDD ablation finds that added components reduce MAE and work compatibly, with combined modules and adversarial loss improving performance. The discussion links this combination to coverage of dishwasher power modes and the branch-wise gates.

C. Ablation Study and Discussion

  • Each added component reduces MAE, and combining components can lower it further; modules combined with adversarial loss greatly improve performance.The ablation evaluates REDD appliances, reporting MAE values averaged over three trials.Source: C. Ablation Study and Discussion · p. 8
  • For fridge, on-state augmentation mainly contributes to improvement by helping adapt to a different test-set power-consumption level.Source: C. Ablation Study and Discussion · p. 8
  • On REDD dishwashers, complete SCANet covers all four training power modes, while SGN and SCANet without branch-wise gates miss modes.The gates were not specifically designed to prevent mode collapse; this is an empirical observation for the dishwasher case.Source: C. Ablation Study and Discussion · p. 8

D. Practicability Verification With Limited Data and Partial Ground Truth

The section tests SCANet against SGN across REDD-to-UK-DALE appliance-state classification using limited training data and partial ground truth. SCANet performs better overall, though microwave transfer is problematic without data from the test household.

D. Practicability Verification With Limited Data and Partial Ground Truth

V. CONCLUSION

SCANet is a scale- and context-aware CNN for NILM that reduces disaggregation error, with adversarial loss and on-state augmentation helping for certain appliances. The conclusion points to combining supervised and unsupervised learning, or using unlabeled or partially labeled data, as future directions.

V. CONCLUSION

  • SCANet significantly reduces appliance-level disaggregation estimation error, and its scale- and context-aware modules contribute to improved accuracy.Source: §V · p. 10
  • Adversarial loss and on-state augmentation are useful for certain appliances.Source: §V · p. 10
  • Future work could combine supervised and unsupervised model outputs or incorporate unlabeled or partially labeled data through semi-supervised learning.Source: §V · p. 10
Loading 1911.07183v1…